bgp multihoming examples - au.int · – router size is related to data rates, not running bgp ......

Advanced Multihoming

BGP Traffic Engineering

1

Service Provider Multihoming

• Previous examples dealt with loadsharing inbound traffic– Of primary concern at Internet edge

– What about outbound traffic?

• Transit ISPs strive to balance traffic flows in both directions– Balance link utilisation

– Try and keep most traffic flows symmetric

– Some edge ISPs try and do this too

• The original Traffic Engineering2


• Balancing outbound traffic requires inbound routing information

– Common solution is “full routing table”

– Rarely necessary

• Why use the “routing mallet” to try solve loadsharing problems?

– “Keep It Simple” is often easier (and $$$ cheaper) than carrying N-copies of the full routing table

3

Service Provider MultihomingMYTHS!!

Common MYTHS

1. You need the full routing table to multihome

– People who sell router memory would like you to believe this

– Only true if you are a transit provider

– Full routing table can be a significant hindrance to multihoming

2. You need a BIG router to multihome

– Router size is related to data rates, not running BGP

– In reality, to multihome, your router needs to:

• Have two interfaces,

• Be able to talk BGP to at least two peers,

• Be able to handle BGP attributes,

• Handle at least one prefix

3. BGP is complex

– In the wrong hands, yes it can be! Keep it Simple!

4

Service Provider Multihoming:Some Strategies

• Take the prefixes you need to aid traffic engineering– Look at NetFlow data for popular sites

• Prefixes originated by your immediate neighbours and their neighbours will do more to aid load balancing than prefixes from ASNs many hops away– Concentrate on local destinations

• Use default routing as much as possible– Or use the full routing table with care

5


• Examples

– One upstream, one local peer

– One upstream, local exchange point

– Two upstreams, one local peer

– Three upstreams, unequal link bandwidths

• Require BGP and a public ASN

• Examples assume that the local network has their own /19 address block

6


One upstream, one local peer

7

One Upstream, One Local Peer

• Very common situation in many regions of the Internet

• Connect to upstream transit provider to see the Internet

• Connect to the local competition so that local traffic stays local

– Saves spending valuable $ on upstream transit costs for local traffic

8


9

AS 110

C

A

Upstream ISP

AS130

Local Peer

AS120


• Announce /19 aggregate on each link

• Accept default route only from upstream

– Either 0.0.0.0/0 or a network which can be used as default

• Accept all routes the local peer originates

10


• Router A Configurationrouter bgp 110

network 121.10.0.0 mask 255.255.224.0

neighbor 122.102.10.2 remote-as 120

neighbor 122.102.10.2 prefix-list my-block out

neighbor 122.102.10.2 prefix-list AS120-peer in

!

ip prefix-list AS120-peer permit 122.5.16.0/19

ip prefix-list AS120-peer permit 121.240.0.0/20

ip prefix-list my-block permit 121.10.0.0/19

!

ip route 121.10.0.0 255.255.224.0 null0 250

11

Prefix filtersinbound


• Router A – Alternative Configurationrouter bgp 110

network 121.10.0.0 mask 255.255.224.0



neighbor 122.102.10.2 filter-list 10 in

!

ip as-path access-list 10 permit ^(120_)+$

!


!

ip route 121.10.0.0 255.255.224.0 null0

12

AS Path filters –more

trusting


• Router C Configurationrouter bgp 110

network 121.10.0.0 mask 255.255.224.0


neighbor 122.102.10.1 prefix-list default in


!


ip prefix-list default permit 0.0.0.0/0

!

ip route 121.10.0.0 255.255.224.0 null0

13


• Two configurations possible for Router A

– Filter-lists assume peer knows what they are doing

– Prefix-list higher maintenance, but safer

– Some ISPs use both

• Local traffic goes to and from local peer, everything else goes to upstream

14

Aside: Configuration Recommendations

• Private Peers– The peering ISPs exchange prefixes they originate

– Sometimes they exchange prefixes from neighbouring ASNs too

• Be aware that the private peer eBGP router should carry only the prefixes you want the private peer to receive– Otherwise they could point a default route to you

and unintentionally transit your backbone

15


One upstream, Local Exchange Point

16

One Upstream, Local Exchange Point

• Very common situation in many regions of the Internet

• Connect to upstream transit provider to see the Internet

• Connect to the local Internet Exchange Point so that local traffic stays local– Saves spending valuable $ on upstream transit

costs for local traffic

• This example is a scaled up version of the previous one

17


18

AS 110

C

A

Upstream ISP

AS130

IXP


• Announce /19 aggregate to every neighbouring AS

• Accept default route only from upstream


• Accept all routes originated by IXP peers

19


• Router A Configurationinterface fastethernet 0/0

description Exchange Point LAN

ip address 120.5.10.1 mask 255.255.255.224

!

router bgp 110

neighbor ixp-peers peer-group

neighbor ixp-peers prefix-list my-block out

neighbor ixp-peers remove-private-AS

neighbor ixp-peers send-community

neighbor ixp-peers route-map set-local-pref in

…next slide

20



neighbor 120.5.10.2 peer-group ixp-peers

neighbor 120.5.10.2 prefix-list peer100 in










...next slide

21


!


ip prefix-list peer100 permit 122.0.0.0/19




!

route-map set-local-pref permit 10

set local-preference 150

!

22

One Upstream, Local Exchange

• Note that Router A does not generate the aggregate for AS110– If Router A becomes disconnected from backbone, then the aggregate

is no longer announced to the IX

– BGP failover works as expected

• Note the inbound route-map which sets the local preference higher than the default– This is a visual reminder that BGP Best Path for local traffic will be

across the IXP

23



network 121.10.0.0 mask 255.255.224.0




!



!

ip route 121.10.0.0 255.255.224.0 null0

24


• Note Router A configuration

– Prefix-list higher maintenance, but safer

– No generation of AS110 aggregate

• IXP traffic goes to and from local IXP, everything else goes to upstream

25

Aside: IXP Configuration

Recommendations• IXP peers

– The peering ISPs at the IXP exchange prefixes they originate

– Sometimes they exchange prefixes from neighbouring ASNs too

• Be aware that the IXP border router should carry only the prefixes you want the IXP peers to receive and the destinations you want them to be able to reach

– Otherwise they could point a default route to you and unintentionally transit your backbone

• If IXP router is at IX, and distant from your backbone

– Don’t originate your address block at your IXP router

26


Two upstreams, one local peer

27

Two Upstreams, One Local Peer

• Connect to both upstream transit providers to see the Internet

– Provides external redundancy and diversity – the reason to multihome

• Connect to the local peer so that local traffic stays local

– Saves spending valuable $ on upstream transit costs for local traffic

28


29

AS 110

C

A

Upstream ISP

AS140

Local Peer

AS120 D

Upstream ISP

AS130


• Announce /19 aggregate on each link

• Accept default route only from upstreams


• Accept all routes originated by local peer

• Note separation of Router C and D

– Single edge router means no redundancy

• Router A

– Same routing configuration as in example with one upstream and one local peer 30



network 121.10.0.0 mask 255.255.224.0




!



!

ip route 121.10.0.0 255.255.224.0 null0

31


• Router D Configurationrouter bgp 110

network 121.10.0.0 mask 255.255.224.0




!



!

ip route 121.10.0.0 255.255.224.0 null0

32


• This is the simple configuration for Router C and D

• Traffic out to the two upstreams will take nearest exit

– Inexpensive routers required

– This is not useful in practice especially for international links

– Loadsharing needs to be better

33


• Better configuration options:

– Accept full routing from both upstreams

• Expensive & unnecessary!

– Accept default from one upstream and some routes from the other upstream

• The way to go!

34

Two Upstreams, One Local PeerFull Routes


network 121.10.0.0 mask 255.255.224.0


neighbor 122.102.10.1 prefix-list rfc1918-deny in


neighbor 122.102.10.1 route-map AS130-loadshare in

!


! See www.cymru.com/Documents/bogon-list.html

! ...for “RFC1918 and friends” list

...next slide

35

Allow all prefixes in apart from RFC1918 and friends


ip route 121.10.0.0 255.255.224.0 null0

!


ip as-path access-list 10 permit ^(130_)+_[0-9]+$

!

route-map AS130-loadshare permit 10

match ip as-path 10


!

route-map AS130-loadshare permit 20


!

36



network 121.10.0.0 mask 255.255.224.0


neighbor 122.102.10.5 prefix-list rfc1918-deny in


!


! See www.cymru.com/Documents/bogon-list.html

! ...for “RFC1918 and friends” list

37

Allow all prefixes in apart from RFC1918 and friends


• Router C configuration:

– Accept full routes from AS130

– Tag prefixes originated by AS130 and AS130’s neighbouring ASes with local preference 120

• Traffic to those ASes will go over AS130 link

– Remaining prefixes tagged with local preference of 80

• Traffic to other all other ASes will go over the link to AS140

• Router D configuration same as Router C without the route-map 38


• Full routes from upstreams

– Expensive – needs lots of memory and CPU

– Need to play preference games

– Previous example is only an example – real life will need improved fine-tuning!

– Previous example doesn’t consider inbound traffic – see earlier in presentation for examples

39

Two Upstreams, One Local PeerPartial Routes: Strategy

• Ask one upstream for a default route

– Easy to originate default towards a BGP neighbour

• Ask other upstream for a full routing table

– Then filter this routing table based on neighbouring ASN

– E.g. want traffic to their neighbours to go over the link to that ASN

– Most of what upstream sends is thrown away

– Easier than asking the upstream to set up custom BGP filters for you

40

Two Upstreams, One Local PeerPartial Routes


network 121.10.0.0 mask 255.255.224.0


neighbor 122.102.10.1 prefix-list rfc1918-nodef-deny in



neighbor 122.102.10.1 route-map tag-default-low in

!

...next slide

41

Allow all prefixes and default in; deny RFC1918 and friends

AS filter list filters prefixes based on origin ASN




!

ip route 121.10.0.0 255.255.224.0 null0

!


ip as-path access-list 10 permit ^(130_)+_[0-9]+$

!

route-map tag-default-low permit 10

match ip address prefix-list default


!

route-map tag-default-low permit 20

!

42



network 121.10.0.0 mask 255.255.224.0




!



!

ip route 121.10.0.0 255.255.224.0 null0

43


• Router C configuration:

– Accept full routes from AS130

• (or get them to send less)

– Filter ASNs so only AS130 and AS130’s neighbouring ASes are accepted

– Allow default, and set it to local preference 80

– Traffic to those ASes will go over AS130 link

– Traffic to other all other ASes will go over the link to AS140

– If AS140 link fails, backup via AS130 – and vice-versa

44


• Router C IGP Configurationrouter ospf 110

default-information originate metric 30

passive-interface Serial 0/0

!

ip route 0.0.0.0 0.0.0.0 serial 0/0 254

• Router D IGP Configurationrouter ospf 110

default-information originate metric 10

passive-interface Serial 0/0

!

ip route 0.0.0.0 0.0.0.0 serial 0/0 254


• Partial routes from upstreams

– Use OSPF to determine outbound path

– Router D default has metric 10 – primary outbound path

– Router C default has metric 30 – backup outbound path

– Serial interface goes down, static default is removed from routing table, OSPF default withdrawn


• Partial routes from upstreams

– Not expensive – only carry the routes necessary for loadsharing

– Need to filter on AS paths

– Previous example is only an example – real life will need improved fine-tuning!

– Previous example doesn’t consider inbound traffic – see earlier in presentation for examples

47

Aside: Configuration Recommendation

• When distributing internal default by iBGP or OSPF/ISIS

– Make sure that routers connecting to private peers or to IXPs do NOT carry the default route

– Otherwise they could point a default route to you and unintentionally transit your backbone

– Simple fix for Private Peer/IXP routers:

ip route 0.0.0.0 0.0.0.0 null0

48


Three upstreams, unequal bandwidths

49

Three upstreams, unequal bandwidths

• Autonomous System has three upstreams– 16Mbps to ISP A

– 8Mbps to ISP B

– 4Mbps to ISP C

• What is the strategy here?– One option is full table from each

• 3x 400k prefixes 1200k paths

– Other option is partial table and defaults from each• How??

50

Strategy

• Two external routers (gives router redundancy)– Do NOT need three routers for this

• Connect biggest bandwidth to one router– Most of inbound and outbound traffic will go here

• Connect the other two links to the second router– Provides maximum backup capacity if primary link fails

• Use the biggest link as default– Most of the inbound and outbound traffic will go here

• Do the traffic engineering on the two smaller links– Focus on regional traffic needs

51

Diagram

• Router A has 16Mbps circuit to ISP A

• Router B has 8Mbps and 4Mbps circuits to ISPs B&C

52

AS 100

B

ISP A

AS110

ISP C

AS130A

ISP B

AS120

Outbound load-balancing strategy

• Available BGP feeds from Transit providers:– Full table

– Customer prefixes and default

– Default Route

• These are the common options on Internet today– Very rare for any provider to offer anything

different

– Very rare for any provider to customise BGP feed for a customer

53


• Accept only a default route from the provider with the largestconnectivity, ISP A– Because most of the traffic is going to use this link

• If ISP A won t provide a default:– Still run BGP with them, but discard all prefixes

– Point static default route to the upstream link

– Distribute the default in the IGP

• Request the full table from ISP B & C– Most of this will be thrown away

– ( Default plus customers is not enough)

54


• How to decide what to keep and what to discard from ISPs B & C?– Most traffic will use ISP A link — so we need to find a good/useful

subset

• Discard prefixes transiting the global transit ISPs– Global transit ISPs generally appear in most non-local or regional AS-

PATHs

• Discard prefixes with ISP A s ASN in the path– Makes more sense for traffic to those destinations to go via the link to

ISP A

55


• Global Transit ISPs include:

209 CenturyLink 3549 Level 3

701 VerizonBusiness3356 Level 3

1239 Sprint 3561 Savvis

1668 AOL TDN 7018 AT&T

2914 NTT America

56

ISP B peering Inbound AS-PATH filter

ip as-path access-list 1 deny _209_








!

ip as-path access-list 1 deny _ISPA_

ip as-path access-list 1 deny _ISPC_

!

ip as-path access-list 1 permit _ISPB$

ip as-path access-list 1 permit _ISPB_[0-9]+$

ip as-path access-list 1 permit _ISPB_[0-9]+_[0-9]+$

ip as-path access-list 1 permit _ISPB_[0-9]+_[0-9]+_[0-9]+$

ip as-path access-list 1 deny .*57

Don t need ISPA and

ISPC prefixes via ISPB

Outbound load-balancing strategy:ISP B peering configuration

• Part 1: Dropping Global Transit ISP prefixes– This can be fine-tuned if traffic volume is not sufficient

– (More prefixes in = more traffic out)

• Part 2: Dropping prefixes transiting ISP A & C network

• Part 3: Permitting prefixes from ISP B, their BGP neighbours, and their neighbours, and their neighbours– More AS_PATH permit clauses, the more prefixes allowed in, the more

egress traffic

– Too many prefixes in will mean more outbound traffic than the link to ISP B can handle

58


• Similar AS-PATH filter can be built for the ISP C BGP peering

• If the same prefixes are heard from both ISP B and C, then establish proximity of their origin ASN to ISP B or C

– e.g. ISP B might be in Japan, with the neighbouring ASN in Europe, yet ISP C might be in Europe

– Transit to the ASN via ISP C makes more sense in this case

59

Inbound load-balancing strategy

• The largest outbound link should announce just the aggregate

• The other links should announce:a) The aggregate with AS-PATH prepend

b) Subprefixes of the aggregate, chosen according to traffic volumes to those subprefixes, and according to the services on those subprefixes

• Example:– Link to ISP B could be used just for Broadband/Dial customers — so

number all such customers out of one contiguous subprefix

– Link to ISP C could be used just for commercial leased line customers — so number all such customers out of one contiguous subprefix

60

Router A: eBGP Configuration Example

router bgp 100

network 100.10.0.0 mask 255.255.224.0

neighbor 122.102.10.1 remote 110


neighbor 122.102.10.1 prefix-list aggregate out

!


ip prefix-list aggregate permit 100.10.0.0/19

!

61

Router B: eBGP Configuration Example

router bgp 100

network 100.10.0.0 mask 255.255.224.0



neighbor 120.103.1.1 prefix-list ISP-B out

neighbor 120.103.1.1 route-map to-ISP-B out



neighbor 121.105.2.1 prefix-list ISP-C out

neighbor 121.105.2.1 route-map to-ISP-C out

!

ip prefix-list aggregate permit 100.10.0.0/19

!

...next slide

62

Router B: eBGP Configuration Example

ip prefix-list ISP-B permit 100.10.0.0/19

ip prefix-list ISP-B permit 100.10.0.0/21

!

ip prefix-list ISP-C permit 100.10.0.0/19

ip prefix-list ISP-C permit 100.10.28.0/22

!

route-map to-ISP-B permit 10

match ip address prefix-list aggregate

set as-path prepend 100

!

route-map to-ISP-B permit 20

!

route-map to-ISP-C permit 10

match ip address prefix-list aggregate

set as-path prepend 100 100

!

route-map to-ISP-C permit 20 63

/21 to ISP Bdial

customers

e.g. Single prepend on ISP B link

/22 to ISP Cbiz

customers

e.g. Dual prepend on ISP C link

What about outbound backup?

• We have:– Default route from ISP A by eBGP

– Mostly discarded full table from ISPs B&C

• Strategy:– Originate default route by OSPF on Router A (with metric 10) — link to

ISP A

– Originate default route by OSPF on Router B (with metric 30) — links to ISPs B & C

– Plus on Router B:

• Static default route to ISP B with distance 240

• Static default route to ISP C with distance 245

– When link goes down, static route is withdrawn

64

Outbound backup: steady state

• Steady state (all links up and active):

– Default route is to Router A — OSPF metric 10

– (Because default learned by eBGP default is in RIB OSPF will originate default)

– Backup default is to Router B — OSPF metric 20

– eBGP prefixes learned from upstreams distributed by iBGP throughout backbone

– (Default can be filtered in iBGP to avoid RIB failure error )

65

Outbound backup: failure examples

• Link to ISP A down, to ISPs B&C up:– Default route is to Router B — OSPF metric 20

– (eBGP default gone from RIB, so OSPF on Router A withdraws the default)

• Above is true if link to B or C is down as well

• Link to ISPs B & C down, link to ISP A is up:– Default route is to Router A — OSPF metric 10

– (static defaults on Router B removed from RIB, so OSPF on Router B withdraws the default)

66

Other considerations

• Default route should not be propagated to devices terminating non-transit peers and customers

• Rarely any need to carry default in iBGP

– Best to filter out default in iBGP mesh peerings

• Still carry other eBGP prefixes across iBGP mesh

– Otherwise routers will follow default route rules resulting in suboptimal traffic flow

– Not a big issue because not carrying full table 67

Router A: iBGP Configuration Example

router bgp 100

network 100.10.0.0 mask 255.255.224.0

neighbor ibgp-peers peer-group

neighbor ibgp-peers remote-as 100

neighbor ibgp-peers prefix-list ibgp-filter out

neighbor 100.10.0.2 peer-group ibgp-peers

neighbor 100.10.0.3 peer-group ibgp-peers

!

ip prefix-list ibgp-filter deny 0.0.0.0/0

ip prefix-list ibgp-filter permit 0.0.0.0/0 le 32

!

68

Three upstreams, unequal bandwidths:

Summary• Example based on many deployed working

multihoming/loadbalancing topologies

• Many variations possible — this one is:

– Easy to tune

– Light on border router resources

– Light on backbone router infrastructure

– Sparse BGP table faster convergence

69

Acknowledgement and Attribution

This presentation contains content and information

originally developed and maintained by the following

organisation(s)/individual(s) and provided for the

African Union AXIS Project

Philip Smith: - [email protected]

Cisco ISP/IXP Workshops

www.apnic.net

Advanced Multihoming

End

71